A Multi-Scale Spatiotemporal Prediction Method for Power Generation of Coastal Wind Turbine Clusters

Abstract

This invention discloses a multi-scale spatiotemporal prediction method for the power generation of coastal wind turbine clusters, mainly including: employing statistical downscaling techniques from numerical weather prediction to perform spatial interpolation and error correction on the target wind turbine locations, generating a high-resolution wind speed prediction sequence covering both short-term and long-term conditions, aligning the predicted wind speed with the measured wind speed, environmental factors, and turbine operating variables, and concatenating them into dynamic input features for each node; and then applying this method to the coastal wind turbine clusters. N The latitude, longitude, and time-series monitoring data of typhoon turbines are encoded into graph nodes. Edge weights are determined based on geographical distance and wake coupling, forming a wind turbine graph network containing both static and dynamic features. The static and dynamic feature sequences are input into a graph neural network containing a spatial attention layer, a temporal recursion layer, and a physical constraint regularization term. Model training is then completed, outputting predicted power generation for each turbine at multiple future time steps and the total power of the wind turbine cluster. This method can provide strong support for wind farm operation scheduling, grid connection management, and renewable energy consumption.

Description

Technical Field

[0001] This invention relates to the field of wind power forecasting and intelligent operation and maintenance technology, and in particular to a multi-scale spatiotemporal prediction method for the power generation of coastal wind turbine clusters. Background Technology

[0002] As the global energy structure accelerates its transformation towards cleaner and lower-carbon energy, the installed capacity of coastal wind farms continues to climb, becoming a crucial renewable energy pillar of the power system. However, the output power of wind turbine clusters is highly random and volatile, influenced by multiple factors such as atmospheric boundary layer evolution, spatial inhomogeneity of wind farms, and turbine operating conditions. Without accurate and robust power forecasting, the power grid must allocate additional reserve capacity and frequently start and stop conventional turbines, increasing dispatch costs and exacerbating wind curtailment and impacting grid frequency stability. High-precision forecasting of wind farm power generation across multiple time scales can provide crucial data for grid economic dispatch, energy storage configuration, demand-side response, and market pricing, playing a vital role in ensuring large-scale integration of renewable energy and the safe and stable operation of the system.

[0003] Existing wind power forecasting methods are mainly based on three categories: physical models, traditional machine learning models, and deep learning models. Physical models rely on numerical weather prediction (NWP), power curves, and wake models, but they struggle to accurately depict the local micro-meteorological characteristics of the land-sea boundary layer, leading to the accumulation of prediction errors over time and space. Traditional machine learning models (such as autoregressive ensemble moving average models and support vector machines) are mostly based on single-unit time series and cannot reveal the spatial coupling between multiple wind turbines. Deep learning models (such as convolutional neural networks, long short-term memory networks, and their composite models) improve time series modeling capabilities, but they typically only extract local or global features, lacking the integration of physical mechanisms such as wind turbine cluster topology, prevailing wind direction, and wake, resulting in limited generalization performance. Furthermore, most methods use simple interpolation for NWP data without incorporating statistical downscaling corrections based on observational data, leading to insufficient prediction resolution. Wind turbine cluster power exhibits significant Kriging-wake correlation spatially and a significant mixture of stationary and non-stationary characteristics temporally. Ignoring this spatiotemporal coupling information makes it difficult to capture the true evolution of wind speed propagation and power fluctuations. Therefore, constructing a spatiotemporal prediction method that can simultaneously mine the spatial dependence between wind turbines, characterize dynamic features at multiple time scales, and embed physical constraints is a key approach to improving the accuracy and reliability of power prediction for coastal wind turbine clusters. Summary of the Invention

[0004] Technical Problem: The purpose of this invention is to provide a multi-scale spatiotemporal prediction method for the power generation of coastal wind turbine clusters. By integrating numerical weather prediction (NWP), measured wind speed, and graph neural network models, this method achieves high-precision prediction of wind turbine cluster power, thereby improving the efficiency and reliability of wind farm operation optimization and grid dispatch.

[0005] Technical Solution: To achieve the above objectives, the present invention provides a multi-scale spatiotemporal prediction method for the power generation of coastal wind turbine clusters, comprising the following steps:

[0006] Step 1, Statistical downscaling wind speed prediction: Spatial interpolation and error correction are performed on key variables of numerical weather prediction (NWP) to output a high-resolution wind speed prediction sequence covering both short-term and long-term conditions; the predicted wind speed is aligned with the measured wind speed and environmental and unit operation variables and spliced ​​into a dynamic input sequence.

[0007] Step 2, Spatiotemporal Graph Construction of Wind Turbine Cluster: Establish nodes based on the latitude and longitude of N wind turbines; determine edge weights based on the geographical distance between wind turbines and the angle with the prevailing wind direction to obtain a weighted undirected graph; use the dynamic input sequence from Step 1 as the dynamic features of each node, and combine it with static features to form a spatiotemporal graph dataset of the wind turbine cluster.

[0008] Step 3, Physics-guided spatiotemporal graph neural network prediction: Input the spatiotemporal graph dataset from Step 2 into a graph neural network containing a spatial attention layer, a temporal recursion layer, and a physical constraint regularization term, and complete the network training to output the predicted power generation of each wind turbine at multiple future time steps and the predicted total power of the wind turbine group.

[0009] in,

[0010] In step 1, the statistical downscaling wind speed prediction step includes:

[0011] Meteorological elements such as wind speed, wind direction, surface temperature, and temperature and wind speed of the pressure layer adjacent to the hub height at the near-surface reference height are extracted from the numerical weather forecast grid data corresponding to the target wind turbine location to form the input vector;

[0012] A sliding time window is used to divide the historical sequence into a short-term window and a long-term window, and a short-term training set D is constructed for each. s With long-term training set ;

[0013] For D s Training a first-order Bayesian dynamic linear model:

[0014]

[0015] right Training a second-order Bayesian dynamic linear model:

[0016]

[0017] in, These are the target variables at time t for the short-term and long-term windows, respectively. These are the vectors of numerical weather prediction (NWP) variables after feature filtering within the short-term and long-term windows, respectively. For basis function mapping; These are the state parameter vectors for the short-term and long-term windows in dynamic regression, respectively. These represent the observation noise for short-term and long-term windows, respectively. These are the short-term and long-term window state evolution noises, respectively. These are the observation noise covariances for the short-term and long-term windows, respectively, describing the measurement error between the numerical weather prediction (NWP) wind speed and the actual observation. These represent the noise covariances of the short-term and long-term window states, respectively, describing the model parameters w. t Uncertainty of random drift over time; N is the number of wind turbines;

[0018] Iterative estimation using the Kalman filter-Expectation Maximization algorithm ; Generate short-term prediction sequences respectively With long-term prediction sequence , and serve as the input feature for the subsequent power prediction model; i is the wind turbine index, i.e., the i-th wind turbine.

[0019] The short-term window ranges from minutes to 6 hours, with a length of h. s The long-term window is 6 hours to 48 hours, with a length of h. l .

[0020] The N wind turbines in step 2 are represented by an undirected graph G=(V, ε):

[0021]

[0022] Where V is the set of nodes; ε is the set of edges, representing the wind turbine pairs in the graph that are considered to have interactions; w ij It is the weight of a single edge; d ij d is the horizontal distance between any two wind turbines; d0 is the distance attenuation scale parameter. The prevailing wind direction; The direction of the line connecting any two wind turbines is... difference;

[0023] The time-invariant geometric parameters and fixed attributes of each wind turbine are organized into a static feature matrix:

[0024]

[0025] Among them, F s For static feature dimensions, Representing the real number field, the time-varying quantities are organized into a dynamic characteristic tensor:

[0026]

[0027] Where T is the length of the history window; Fd For dynamic feature dimensions.

[0028] The static characteristics include: wind turbine latitude and longitude coordinates, hub height and impeller diameter, rated power, rated speed, unit model and unit type, cut-in wind speed, rated wind speed, cut-out wind speed, hub azimuth angle, control system type and its main control strategy parameters.

[0029] The dynamic characteristics include: the horizontal average wind speed, gust wind speed and direction, wind speed measured by the anemometer, instantaneous gust wind speed, wind direction and turbulence intensity, measured air density, ambient temperature, humidity and air pressure; generator power, reactive power, generator speed, pitch angle, yaw angle and yaw error, active pitch coefficient and power limit parameters.

[0030] In step three, the spatial attention layer is modeled by an n-layer improved graph attention network (GAT), and the attention weights and node updates of each layer satisfy the following:

[0031]

[0032] Where, α ij Let be the attention coefficient of edge (i, j), which refers to the relative contribution of information from neighbor node j to the weighted aggregation during a single update of node i in the spatial attention layer, satisfying the following condition: ; This represents the input feature vector of node i; The updated feature vector for node i; The weight matrix is ​​a linear transformation matrix; This is the attention weight vector; x is a non-linear activation function; j Let be the input feature vector of node j.

[0033] The dynamic characteristics of the nodes, modeled using gated recurrent units (GRUs), satisfy the following:

[0034]

[0035] in, To fuse feature vectors in space at time step t; r t To reset the gate, control the proportion of forgotten old information; z t To update the door, control the proportion of the old state that is retained; This is the candidate hidden state; The hidden states of the GRU at the previous and current time steps; This is element-wise multiplication; The input weight matrix; Let H be the recursive weight matrix.

[0036] The static and dynamic features described in step three are fused element-by-element and then directly linearly mapped to obtain the power prediction, satisfying the following:

[0037]

[0038] in, This is a static embedding; s is a static feature vector; These are static feature mapping parameters; f is the dynamic feature vector; f is the fused feature. To predict head weights and biases; This is the predicted power generation value for a single wind turbine, for all wind turbines. Summing yields the predicted total wind power; dimension H; This is element-wise multiplication.

[0039] The model training described in step three, taking into account both measurement errors and physical constraints, satisfies the following:

[0040]

[0041] in, P represents the predicted power output of a single wind turbine, and P represents the measured power output of the wind turbine. curve The output is the standard power curve; (i, j) represents the adjacent fan pair; P wake (i, j) represents the coupling power calculated based on the momentum conservation wake model; For physical regularization weights.

[0042] Beneficial effects: After adopting the above solution, the advantages of the present invention are as follows:

[0043] (1) By statistical downscaling of numerical weather prediction and spatiotemporal consistency calibration of anemometer measured data, high-quality input features are constructed, which significantly reduces the impact of wind speed extrapolation error and data gap on power prediction.

[0044] (2) By introducing graph attention network and temporal GRU structure, the spatial synergy effect of wind turbine group and dynamic changes at multiple time scales are captured simultaneously, which greatly improves the accuracy of power generation prediction.

[0045] (3) Embed power curve consistency and wake energy conservation regularization terms in the loss function to avoid non-physical outputs from deep models and enhance the interpretability and generalization ability of prediction results.

[0046] (4) High-resolution, low-error power prediction provides a reliable basis for active power smoothing control, energy storage dispatch and power trading in wind farms, reducing standby costs and improving wind power absorption rate. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the grid nodes and measurement point locations for numerical weather prediction;

[0048] Figure 2 This is a schematic diagram of the spatiotemporal nodes of the wind turbine cluster;

[0049] Figure 3 It is a spatiotemporal graph neural network structure diagram. Detailed Implementation

[0050] The technical solution and beneficial effects of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] The multi-scale spatiotemporal prediction method for the power generation of coastal wind turbine clusters of the present invention includes the following steps:

[0052] Step 1, Statistical downscaling wind speed prediction: Spatial interpolation and error correction are performed on key variables of numerical weather prediction (NWP) to output a high-resolution wind speed prediction sequence covering both short-term and long-term conditions; the predicted wind speed is aligned with the measured wind speed and environmental and unit operation variables and spliced ​​into a dynamic input sequence.

[0053] Step 2, Spatiotemporal Graph Construction of Wind Turbine Cluster: Establish nodes based on the latitude and longitude of N wind turbines; determine edge weights based on the geographical distance between wind turbines and the angle with the prevailing wind direction to obtain a weighted undirected graph; use the dynamic input sequence from Step 1 as the dynamic features of each node, and combine it with static features to form a spatiotemporal graph dataset of the wind turbine cluster.

[0054] Step 3, Physics-guided spatiotemporal graph neural network prediction: Input the spatiotemporal graph dataset from Step 2 into a graph neural network containing a spatial attention layer, a temporal recursion layer, and a physical constraint regularization term, and complete the network training to output the predicted power generation of each wind turbine at multiple future time steps and the predicted total power of the wind turbine group.

[0055] In step 1, the NWP statistical downscaling wind speed prediction step includes:

[0056] The input vector is constructed by extracting meteorological elements such as wind speed (gusts and averages) at near-surface reference height, wind direction, surface temperature, and temperature and wind speed of the pressure layer adjacent to the hub height from the numerical weather forecast grid data corresponding to the target wind turbine location.

[0057] The historical series is divided into short-term windows (minutes to 6 hours, length h) using a sliding time window. s ) and long-term windows (6h to 48h, length h) l ), and construct short-term training sets D respectively. s With long-term training set D l .

[0058] For D s Training a first-order Bayesian dynamic linear model:

[0059]

[0060] For D l Training a second-order Bayesian dynamic linear model:

[0061]

[0062] in, These are the target variables at time t for the short-term and long-term windows, respectively. This refers to the NWP predictor variable vector within the corresponding window after feature filtering; For basis function mapping; This is the state parameter vector for dynamic regression; To observe noise; This is state evolution noise; To observe the noise covariance, describe the measurement error between NWP wind speed and actual observation; Let w be the state evolution noise covariance, describing the model parameters w. t Uncertainty of random drift over time.

[0063] Iterative estimation using Kalman filtering and expectation-maximization algorithms Short-term prediction sequences are generated respectively. With long-term prediction sequence , and serve as input features for subsequent power prediction models.

[0064] like Figure 2 As shown, in step 2, the N wind turbines are represented as an undirected graph G=(V, ε):

[0065]

[0066] Where, d ij The horizontal distance between any two wind turbines; The prevailing wind direction; The direction of the line connecting any two wind turbines is... difference. Based on at least 6 months of wind direction data (actual measurement or retrospective calculation), the mean value is calculated using circular mean; it can be updated monthly / quarterly.

[0067] Determine the edge set ε using the radius threshold method:

[0068]

[0069] Where D is the impeller diameter, and β∈[3, 7].

[0070] The time-invariant geometric parameters and fixed attributes of each wind turbine are organized into a static feature matrix:

[0071]

[0072] Among them, F s The static feature dimension includes: wind turbine latitude and longitude coordinates; hub height (80m-160m) and rotor diameter (100m-200m); rated power (2MW-12MW), rated speed, turbine model and type (fixed pitch variable speed, variable pitch variable speed, etc.); cut-in / rated / cut-out wind speed (3–4m / s / 10–14m / s / 20–25m / s); hub azimuth (initial turbine orientation); control system type (e.g., fixed pitch variable speed, variable pitch variable speed) and its main control strategy parameters.

[0073] Organize the time-varying quantities into a dynamic feature tensor:

[0074]

[0075] Where T is the length of the history window; F d For dynamic features, the following are included: NWP horizontal average wind speed, gust wind speed and direction; anemometer-measured horizontal wind speed, instantaneous gust wind speed, wind direction and turbulence intensity; measured air density, ambient temperature, humidity, and air pressure; generator power, reactive power, generator speed, pitch angle, yaw angle and yaw error; active pitch coefficient and power limit parameters. Data missing percentage ≤10%; single-point missing data is filled using linear interpolation or forward filling (≤3 sampling intervals); missing segments >30 minutes (or 3 intervals) are merged into "invalid segments".

[0076] like Figure 3 As shown, the spatial dependency between wind turbines is modeled using a two-layer improved graph attention network (GAT). The attention weights and node updates of each layer satisfy the following:

[0077]

[0078] Where is the attention coefficient of edge (i, j), satisfying ; This represents the input feature vector of node i; The updated feature vector for node i; The weight matrix is ​​a linear transformation matrix; This is the attention weight vector; It is a non-linear activation function.

[0079] The time dependence of wind speed and power sequences is modeled using gated cyclic units (GRUs), satisfying the following:

[0080]

[0081] in, To fuse feature vectors in space at time step t; Let H be the GRU hidden state at the previous time step and the current time step; r be the dimension of the hidden state. t To reset the gate, control the proportion of forgotten old information; z t To update the door, control the proportion of the old state that is retained; This is the candidate hidden state; This is element-wise multiplication; The input weight matrix; This is the recursive weight matrix.

[0082] Power prediction is obtained by directly linearly mapping static and dynamic features after element-wise fusion, satisfying the following conditions:

[0083]

[0084] Where s is the static feature vector; These are static feature mapping parameters; It is a static embedding; f is the dynamic feature vector; f is the fused feature. To predict head weights and biases; The predicted power output of a single wind turbine is given by summing the values ​​for all wind turbines to obtain the predicted total power output of the wind farm.

[0085] The training objective function considers both measurement error and physical constraints, satisfying the following:

[0086]

[0087] Where P is the measured power of the wind turbine; P curve The output is the standard power curve; (i, j) represents the adjacent wind turbine pairs in the figure; P wake (i, j) represents the coupling power calculated based on the momentum conservation wake model; For physical regularization weights.

Claims

1. A multi-scale spatiotemporal prediction method for the power generation of coastal wind turbine clusters, characterized in that... Includes the following steps: Step 1, Statistical downscaling wind speed prediction: Spatial interpolation and error correction are performed on key variables of numerical weather prediction (NWP) to output a high-resolution wind speed prediction sequence covering both short-term and long-term conditions; the predicted wind speed is aligned with the measured wind speed and environmental and unit operation variables and spliced ​​into a dynamic input sequence. Step 2, Spatiotemporal Graph Construction of Wind Turbine Cluster: Establish nodes based on the latitude and longitude of N wind turbines; determine edge weights based on the geographical distance between wind turbines and the angle with the prevailing wind direction to obtain a weighted undirected graph; use the dynamic input sequence from Step 1 as the dynamic features of each node, and combine it with static features to form a spatiotemporal graph dataset of the wind turbine cluster. Step 3, Physics-guided spatiotemporal graph neural network prediction: Input the spatiotemporal graph dataset from Step 2 into a graph neural network containing a spatial attention layer, a temporal recursion layer, and a physical constraint regularization term, and complete the network training to output the predicted power generation of each wind turbine at multiple future time steps and the total power of the wind turbine group. in, In step one, the statistical downscaling wind speed prediction step includes: Meteorological elements such as wind speed, wind direction, surface temperature, and temperature and wind speed of the pressure layer adjacent to the hub height at the near-surface reference height are extracted from the numerical weather forecast grid data corresponding to the target wind turbine location to form the input vector; A sliding time window is used to divide the historical sequence into a short-term window and a long-term window, and short-term training sets are constructed for each. With long-term training set ; right Training a first-order Bayesian dynamic linear model: ， right Training a second-order Bayesian dynamic linear model: ， in, , These are the target variables at time t for the short-term and long-term windows, respectively. , These are the vectors of numerical weather prediction (NWP) variables after feature filtering within the short-term and long-term windows, respectively. For basis function mapping; , These are the state parameter vectors for the short-term and long-term windows in dynamic regression, respectively. , These represent the observation noise for short-term and long-term windows, respectively. , These are the short-term and long-term window state evolution noises, respectively. , These are the observation noise covariances for the short-term and long-term windows, respectively, describing the measurement error between the numerical weather prediction (NWP) wind speed and the actual observation. , These represent the noise covariances of the short-term and long-term window states, respectively, describing the model parameters w. t Uncertainty of random drift over time; Represents a normal distribution; Iterative estimation using the Kalman filter-Expectation Maximization algorithm , , , ; Generate short-term prediction sequences respectively With long-term prediction sequence , and serve as input features for subsequent power prediction models; It is the wind turbine index, that is, the first... One fan; The N wind turbines in step two are represented as an undirected graph. : ， Where V is the set of nodes; It is a set of edges, representing wind turbine pairs in the graph that are considered to have interactions; It is the weight of a single edge; The horizontal distance between any two wind turbines; It is a distance attenuation scale parameter; The prevailing wind direction; The direction of the line connecting any two wind turbines is parallel to... difference; The time-invariant geometric parameters and fixed attributes of each wind turbine are organized into a static feature matrix: ， in, For static feature dimensions, Representing the real number field, the time-varying quantities are organized into a dynamic characteristic tensor: ， Where T is the length of the history window; The dynamic feature dimension; N is the number of wind turbines; In step three, the spatial attention layer is modeled by an n-layer improved graph attention network (GAT), and the attention weights and node updates of each layer satisfy the following: ， in, For the edge The attention coefficient refers to the attention coefficient of nodes in the spatial attention layer. During one update, from neighboring nodes The relative contribution of information in the weighted aggregation, satisfying ; Represents a node The input feature vector; For nodes Updated feature vector; The weight matrix is ​​a linear transformation matrix; This is the attention weight vector; It is a non-linear activation function; For nodes The input feature vector.

2. The multi-scale spatiotemporal prediction method for power generation of coastal wind turbine clusters as described in claim 1, characterized in that, The short-term window ranges from minutes to 6 hours, with a length of... The long-term window is 6 hours to 48 hours. .

3. The multi-scale spatiotemporal prediction method for power generation of coastal wind turbine clusters as described in claim 1, characterized in that, The static characteristics include: wind turbine latitude and longitude coordinates, hub height and impeller diameter, rated power, rated speed, unit model and unit type, cut-in wind speed, rated wind speed, cut-out wind speed, hub azimuth angle, control system type and its main control strategy parameters.

4. The multi-scale spatiotemporal prediction method for power generation of coastal wind turbine clusters as described in claim 1, characterized in that, The dynamic characteristics include: the horizontal average wind speed, gust wind speed and direction, wind speed measured by the anemometer, instantaneous gust wind speed, wind direction and turbulence intensity, measured air density, ambient temperature, humidity and air pressure; generator power, reactive power, generator speed, pitch angle, yaw angle and yaw error, active pitch coefficient and power limit parameters.

5. The multi-scale spatiotemporal prediction method for power generation of coastal wind turbine clusters as described in claim 1, characterized in that, The dynamic characteristics of the nodes, modeled using gated recurrent units (GRUs), satisfy the following: in, To fuse feature vectors in space at time step t; To reset the gate, control the proportion of forgotten old information; To update the door, control the proportion of the old state that is retained; This is the candidate hidden state; The hidden states of the GRU at the previous and current time steps; This is element-wise multiplication; The input weight matrix; Let H be the recursive weight matrix.

6. The multi-scale spatiotemporal prediction method for power generation of coastal wind turbine clusters as described in claim 1, characterized in that, The static and dynamic features described in step three are fused element-by-element and then directly linearly mapped to obtain the power prediction, satisfying the following: ， in, It is a static embedding; These are static feature vectors; These are static feature mapping parameters; It is a dynamic feature vector; Features after fusion; To predict head weights and biases; This is the predicted power generation value for a single wind turbine, for all wind turbines. Summing yields the predicted total wind power; dimension H; This is element-wise multiplication.

7. The multi-scale spatiotemporal prediction method for power generation of coastal wind turbine clusters as described in claim 1, characterized in that, The model training described in step three, taking into account both measurement errors and physical constraints, satisfies the following: ， in, This represents the predicted power generation of a single wind turbine unit. This represents the actual measured power of the fan. Standard power curve output; For adjacent wind turbine pairs; The coupling power is calculated based on the momentum conservation wake model; For physical regularization weights.

Images